Lumilens emerges from stealth mode with more than $700M in new funding

Lumilens, a developer of connectivity systems for AI infrastructure, earlier this month (6th August) emerged from stealth, having been founded “to solve networking bottlenecks limiting AI scale”. Two years since its founding in San Jose, California, the company is now shipping its first product into production AI data centers under a multi-billion dollar customer agreement.

Lumilens designs and manufactures optical interconnects for both types of networks inside an AI data center – the scale-up network that directly connects GPUs together within a single compute system, and the scale-out fabric that stitches different racks and clusters together.

The company has also just announced more than $700 million in new funding co-led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures, and Spark Capital, with participation from Addition, Alkeon, Harbourvest, J.P. Morgan Private Capital, Mayfield, MVP Ventures, Qualcomm Ventures, Peak XV, Redpoint Ventures, Seifdune and other top-tier investors. The financing values the company at $5.51 billion and brings its total capital raised to over $900 million.

The world’s largest technology companies are spending historic sums on AI infrastructure – Citigroup projects the total tech sector’s investment in AI infrastructure will pass $2.8 trillion by 2029 – but buying GPUs is no longer the hard part, says Lumilens; the challenge is connecting them.

‘Millions of transceivers needed’

The challenge is emerging in two critical places inside the AI data center: scale-out and scale-up fabrics. In the scale-out fabric that ties racks and rows of GPUs together, every processor added multiplies the optical transceivers required.

Lumilens stated that “a single 400,000 GPU data center requires more than 2.4 million transceivers and more than five million fiber strands from a market already unable to keep pace. McKinsey projects that 800G optical transceiver production could fall 40-60% short of industry demand through 2027, with 1.6T transceiver shortfalls of 30-40% persisting through 2029.”

Inside the rack, physics has imposed a ceiling on the scale-up fabric, where GPUs are wired directly to one another over copper connections. At the data rates AI demands, electrical signals survive only about a meter and a half over copper, confining a tightly coupled domain to a single rack and hundreds of GPUs. When a hyperscaler roadmap calls for thousands of GPUs, breaking that ceiling means replacing copper with photonics. Lumilens was founded to address that constraint.

In early 2024, the company’s founding team set out to solve the connectivity problems the largest AI clusters foreshadowed – problems a leading hyperscaler was already confronting as it sized the clusters its AI roadmap would require. Within two years, Lumilens’ initial scale-out product finished qualification and then started shipping into that hyperscaler’s production AI data centers under a multi-billion-dollar agreement.

GPU shift

“The constraint on AI has shifted from how many GPUs you can buy to how many you can connect,” said Ankur Singla, founder and CEO of Lumilens. “Whether they are training frontier models or serving them to hundreds of millions of users, hyperscalers told us they need the same two things: far more optical capacity for the networks they run today, and a path to directly connecting thousands of GPUs together into a single cluster. We built Lumilens to deliver both – and to manufacture at the volume, quality, and pace this buildout demands.”

The Lumilens portfolio spans the full connectivity stack of an AI data center:

·  Scale-up silicon and systems (NPO and CPO): near-package and co-packaged optics that bring optical I/O directly to the GPU, enabling thousands of processors — and, on the company’s roadmap, tens of thousands — to operate as a single, tightly coupled domain; and

·  Scale-out pluggable transceivers (800G, 1.6T, and beyond): optical modules that replace copper links across racks and rows, carrying more data at lower power for the largest AI fabrics.

Every product is built on LumiCoreTM, the company’s unified technology platform spanning silicon photonics, mixed-signal ICs, electrical-optical interposers, and optical systems, all developed in-house. The company has designed a novel electrical-optical interposer technology that dramatically simplifies photonic chip assembly, along with proprietary manufacturing process recipes, custom robotics, and test automation – scaling capacity through manufacturing partners and its own large-scale facilities.

Investor and analyst comments

“Over the past several years, the conversation around AI infrastructure has centered on access to compute,” said Umesh Padval, Managing Partner, Seligman Ventures. “Increasingly, the constraint today is shifting to connectivity and how efficiently thousands of GPUs can operate as a single system.”

“The critical component in scaling AI is shifting from compute to connectivity,” said Gavin Baker, Managing Partner at Atreides Management. “Focusing on that core AI challenge, Lumilens has moved from its founding toward a qualified product deployed inside a hyperscaler’s production data centers within two years, and is building the manufacturing engine to scale with demand while bringing much-needed diversification to AI supply chains.”

“Connecting GPUs is the bottleneck now, and copper has hit a wall the industry cannot engineer around,” said Navin Chaddha, Managing Partner at Mayfield. “That’s why we partnered with Ankur Singla for the third time to help build Lumilens from inception, led their seed round, and have participated in every round since.”